长明 · Evenlight

1.2.146 起改名为「长明 · Evenlight」，之前叫 Nightshift · 夜班；数据和自动更新都不受影响。网站与使用文档：https://20chuangge.github.io/nightshift/

你休息时，灯还亮着。一个有自己电脑的 AI 助理：说一句要做什么，它在电脑上把事情办完。

官网和下载：https://20chuangge.github.io/nightshift/

• 一个助理就够：默认的「助理」什么都能做，也能自己建 Bot、拉群分工。
• 多台机器：闲置电脑或云主机填上 IP、账号、密码，自动装好；切换机器，整个界面跟着切。支持 Ubuntu、Rocky、Windows。
• 看得见、能接管：每台机器的屏幕实时可看，随时接管鼠标键盘。
• 本机先问你，远程机自主：在你自己的电脑上，删文件、发消息这类操作会先问；远程机器放手干，危险命令一律拦下。
• 记忆、技能、定时任务：记住你的习惯，把做过的流程存成技能，按时自动跑。
• 附件和语音：拖文件、贴截图，或者按 Ctrl+D 直接说。
• 一处配置：模型和语音密钥配一次，所有机器共用。

下载

到 Releases 下载 Evenlight-Setup-x.y.z.exe（安装版）或 Evenlight-Portable-x.y.z.exe（免安装）。首次打开后在「设置 → 模型」填一个模型密钥（DeepSeek、通义、Kimi、OpenAI 等 OpenAI 兼容接口都行）。

从源码运行

npm install
npm start                # 服务 + 网页界面 http://127.0.0.1:4390
npm run desktop          # 桌面应用
npm run smoke            # 全量自测
npm run desktop:build    # 出 Windows 安装包 → desktop/dist/

需要 Node ≥ 20。添加远程机器、空白机测试方案见 docs/（HUB.md、TEST-PLAN.md）。

许可证

MIT：可以自由使用、修改和再分发，保留版权声明即可。

---

以下是开发参考（架构、接口、扩展方式）。

Quick start

# 1) 起后端 + 浏览器打开
npm start                      # http://127.0.0.1:4390
npm run smoke                  # 1043 assertions, boots the real server on a free port

# 2) 桌面应用（Electron 壳）
npm run desktop                # 开发模式，等同 npm start 但带窗口/托盘
npm run desktop:build          # 出 Windows 安装包 → desktop/dist/

# 3) 手机端
npm run mobile:sync            # 把 public/ 同步进 mobile/www 并 cap sync

PORT=0 npm start picks a free port; the server prints one JSON line with the actual URL and logs to stderr.

Mock mode (default)

No configuration. ASSIST_MODEL_PROVIDER defaults to mock: a deterministic planner that reads keywords out of the last user message and replays a fixed plan of tool calls (fs_list → fs_write → memory_write → task_finish). It is not a language model — it is a fixture that exercises every code path, which is exactly what makes the smoke test meaningful.

Try these messages:

Message	What it exercises
List the files in the workspace and write a note	LOW-risk tools, streaming, memory write
Remember that I prefer metric units	per-bot memory
Write a note outside the workspace (Ops)	HIGH-risk approval card
Hand off to the researcher: summarise welcome.md	bot-to-bot handoff, separate turn
Save this as a skill called inbox-triage (Researcher)	skill library write, auto-approved MEDIUM
Create a routine every 30 minutes	routine creation

Real mode (opt-in)

Any OpenAI-compatible /chat/completions endpoint works:

export ASSIST_MODEL_PROVIDER=openai
export ASSIST_API_KEY=sk-...                      # never committed
export ASSIST_MODEL=gpt-4o-mini                   # optional
export ASSIST_BASE_URL=https://api.openai.com/v1  # optional: any compatible endpoint
npm start

The real client sends the tool schemas, parses tool_calls, and feeds results back exactly like the mock path. Streaming tokens from the provider is not enabled (see "what this replica does NOT do"); the UI still streams the assistant text because the server emits deltas for each completed message.

Environment variables

Variable	Default	Meaning
PORT	4390	HTTP port. 0 = any free port.
HOST	127.0.0.1	Bind address. Loopback by default, on purpose.
ASSIST_HOME	repo root	Relocates the whole installation (used by the smoke test).
ASSIST_DATA_DIR	<home>/data	Sessions, audit, routines, memory.
ASSIST_WORKSPACE_DIR	<home>/workspace	The shared workspace.
ASSIST_MODEL_PROVIDER	mock	mock or openai.
ASSIST_BASE_URL / ASSIST_API_KEY / ASSIST_MODEL	—	Real provider settings.
ASSIST_MODEL_TIMEOUT_MS	120000	Per-request timeout.
CHROME_PATH	unset	Enables the optional browser_open tool.
ALLOW_PRIVATE	unset	1 lets http_fetch/browser_open reach LAN addresses.

---

This is the beginning of the README. Read the rest at the original. Read the full README ↗